Hand-Written-Digit-Recognition

Trained on MNIST dataset

main.py

It uses simple neural neywork with Stochastic Gradient Descent and Backpropogation

cross-entropy-costfuntion.py

1. cross-entropy cost function instead of quadratic cost function L2-regularisation
2. Weights are initiazed to mean zero

The best efficiency I got was 98.87% with a couple of convolutional layers, a fully connected layer of 640 neurons, learning rate of 0.1

Note : I used very small dataset to train, more likely reason for overfiiting